AR-Sim: A High-fidelity 3D Simulator for Autonomous Racecars

This thesis introduces a versatile simulation environment for autonomous racing vehicles, developed at the TUM Autonomous Vehicle System Lab. Initially created for F1Tenth cars, the simulator is adaptable for any vehicle type, including full-scale models. It feature a C++ API and a custom Unity executable for realistic graphics and sensor simulations. A ROS2 extension package is included for easy software-in-the-loop testing, while a Python wrapper implements a Gymnasium environment for machine learning algorithms. The project not only addresses the limitations of the current F1Tenth gym simulator, which is confined to 2D environments and lacks capabilities to include camera and 3D LiDAR sensors, but overcome the lack of open source autonomous racing simulators for vehicles. The existing simulators are few and with limited generalizability, resulting in a trade off between customization of vehicle dynamic modeling and 3D realism. On the other side autonomous road vehicle simulation environments present advanced capabilities but are very complex and demand substantial computational resources, in addition none of them allow fully customization of the vehicle dynamic model employed and mostly only rely on the PhysX physics engine. This new simulator bridges the gap with unique features, offering both PhysX and external physics engines, and provides an easy-to-use API for custom dynamic models. User-friendly YAML configuration files enable seamless setup of the simulation environment. Custom race tracks can be integrated from pre-made 3D models, CSV files, or generated randomly using varied barriers and materials. The simulator also features a LiDAR model and an RGB camera sensor, both of which are fully parameterized to match real sensor specifications. Multiple sensors can be placed on a car, with the option to include more than one of each type. This open-source simulator is a valuable tool for research and education, especially in the F1Tenth community, and its flexibility makes it suitable for a broad spectrum of autonomous racing vehicle applications.